Recurrent Neural Network-Based NMPC for Nonlinear Processes

This paper investigates the application of model predictive control (MPC) based on recurrent neural networks (RNNs) in addressing challenges posed by nonlinear process dynamics. The study considers a continuous-flow stirred tank reactor characterized by complex equilibrium series kinetics; informative data generation, model training, and model testing are discussed herein. A comparative analysis with the nominal scenario, based on the first-principles model with no plant/model mismatch, and with some traditional linear data-driven models, highlights the competitive performance of RNN-based MPC in simulated control scenarios. Suitable key performance indicators demonstrate the effectiveness in controlling optimal targets, tracking setpoint variations, and rejecting disturbances. The proposed RNN-based MPC offers a competitive approach to enhance control strategies in complex dynamic nonlinear systems.